Global_Hack6

Homelessness Garph Analytics

ggplot(ml_data, aes(EntryDateMonth)) + geom_histogram(binwidth = 1)

• As we see, we have sessional pattern here.
• More people comes to the center during summer and winter
• Less people during Spring and fall
• We need more resources during peak sesson to ensure that we can provide the quality services.

ggplot(ml_data, aes(EntryDateMonth, fill = as.factor(Destination))) + geom_histogram(binwidth = 1)

• Higher the factor number means better result after they competed the program.
• We have great result dsuring the summer but not so good result during the winter.
• Let’s see why is that the case.

ggplot(ml_data, aes(EntryDateMonth, fill = as.factor(TypeProvided))) + geom_histogram(binwidth = 1)

• Factor variables are representing specific service they have received during the enrollment.
• We know that during the summer, the program does great job and ‘A2’ services are offered the most during the summer. (A2 = Community service/service learning (CSL))
• During the winter session, we offer less ‘A2/ service and more ’B1’ services. We might need to change the service stratagies. (B1 = Rental assistance)

ggplot(ml_data, aes(EntryDateMonth, fill = as.factor(VADisabilityService))) + geom_histogram(binwidth = 1)

• Most of the clients during the summer receive VA Disability Service which could be the potensial reason why we have such a great outcome during the summer. correlation but not causation.

Heat Map Graph Analytics

g2

## Warning: Removed 69 rows containing non-finite values (stat_density2d).

g1

## Warning: Removed 1110 rows containing missing values (geom_point).

• This map represents the intensity of the homelessness.
• More freqeunt homelessness has bigger diameter of the circle.
• Let see which area need more resources.

print("Top 10 the Most Homelessness Area")

## [1] "Top 10 the Most Homelessness Area"

top10_zip$LastPermanentZIP

##  [1] 63101 63136 63118 63103 63111 63125 63114 63104 63107 63143

• Above zip code is the top 10 zipcodes where the homelessness is frequently occuring.
• According to these information, we can distribute available resources more intelligently for higher quality services.

Predictive modeling

• If we can predice the successfulness of the client prior to the service enrollment, then it is greatly beneficial for the center for having a chance to improve the 1 to 1 service if it predicts the negative outcomes.
• However, the current data sets have too much features which the model does not need them all.
• Let’s plot the most important features to select the best predictor variables.

plotFilterValues(var_imp, feat.type.cols=TRUE)

imp_feat <- (var_imp$data %>% arrange(-information.gain) %>% top_n(7))$name

## Selecting by information.gain

imp_feat

## [1] "EntryDateMonth"      "Discharge_Status.1"  "Discharge_Status.2" 
## [4] "Age"                 "Employed"            "TypeProvided"       
## [7] "VADisabilityService"

• We already analized some of the variables previously and found some significant intuition.
• As long as we have those 7 variables, we can make great predictive model.
• Let’s do some demonstration

head(test, 3)

##   TypeProvided VADisabilityService EntryDateMonth Discharge_Status.1
## 2           B1                   0              1                  0
## 5           B1                   0              1                  0
## 7           B1                   0              1                  0
##   Discharge_Status.2 Age Employed Destination
## 2                  1   4        1           3
## 5                  1   4        1           3
## 7                  1   4        1           3

pred.rf.test <- predict(mdl.rf, test)
conf.mtx <- confusionMatrix(pred.rf.test, test$Destination)
conf.mtx$overall

##       Accuracy          Kappa  AccuracyLower  AccuracyUpper   AccuracyNull 
##      0.8656172      0.7352898      0.8639407      0.8672805      0.6662625 
## AccuracyPValue  McnemarPValue 
##      0.0000000            NaN

• As you can see, we have over 86% of accuracy to determine the successfulness of the particular individual prior to the enrollment without collecting too much data.

Improvement TODO:

• Explore more featurese and perform decent feature engineering process.
• Construct Deep Neuron Network through tensorflow to build more compelx predictive model.
• Use Spark distributed computing system to find the best hyper-parameters for Deep Neuron Network.

Challenges of this project

• Cleaning multiple, exreamly dirty, and small dataset and merging them into descent size was one of the challenges.